Counting Is the Easy Part: Bin Decomposition, Two-Key Putaway, and the Variance a Physical Stock-Take Actually Owes

1.1The three error terms

The product's own material names three failure modes that recur every year, and they map cleanly onto distinct terms in a count's error budget.

1.2Contributions

1. Bin decomposition stated as a conservation constraint, fixing the granularity at which a count is checkable (Section 4.1).
2. Two-key putaway as a control function in Shingo's sense, with its error model (Section 4.2).
3. Tag locking and missed-bin detection as structural eliminations of the two dominant count errors (Section 4.3).
4. Absolute against net variance, and why a reconciled total is not evidence of an accurate record (Section 4.4).
5. An eight-dimension capability reference framework for stock-take systems (Section 6.3).

2.1Trueness and precision are different failures

ISO 5725 separates trueness — the closeness of a mean to a reference value — from precision, the scatter of repeated measurements. The distinction is routinely collapsed in stock-take reporting, and collapsing it is the source of the most common misreading of a variance report.

A count with poor trueness is systematically wrong in one direction: a scanner configured for the wrong pack size, a bin standard that counts crates as pieces. A count with poor precision is scattered: some locations high, some low, no pattern. The first is a configuration defect and is cheap to fix; the second is a process defect and is not.

Section 4.4 shows that net variance measures only the first and absolute variance measures both, which is why reporting one number lets a poor count pass.

2.2Control against warning at the point of error

Shingo distinguishes the control function of an error-proofing device — the next action becomes impossible — from the warning function, where the operator is told and may proceed. The distinction is decisive in a stock-take because the environment is the worst case for warnings: temporary staff, unfamiliar layout, time pressure, and a lamp nobody has been trained to fear.

The putaway design of Section 4.2 is a control function. It does not tell the operator that the container and location disagree; it refuses to store the record until they agree. The distinction is the difference between an error rate that falls and one that is documented.

2.3What makes an object countable

A count presupposes a set of distinguishable objects. GS1 provides the Global Trade Item Number for a trade item class and the Serial Shipping Container Code for an individual logistic unit; ISO/IEC 15459-1 covers transport units outside a GS1 scheme; the General Specifications govern symbology selection and print quality, which decide whether a tag scans at all on the first attempt.

The class-against-instance distinction is what makes bin-level counting possible. A tag identifying only the part number tells a counter what is in front of them; a tag identifying this container tells the system which of forty identical containers has been counted, which is the fact tag locking depends on.

The event model is EPCIS, standardised as ISO/IEC 19987, with its vocabulary in ISO/IEC 19988. A count observation is an event with a business step and a disposition like any other, which means a stock-take integrates into a traceability record rather than sitting beside it as a spreadsheet.

3.1Receipt, decomposition and tagging

A receipt is captured against its document and invoice, with the part, its line-side location, its standard bin quantity and its shelf life in days. Two quantities are recorded: the quantity the invoice claims and the quantity actually received.

The two-quantity pattern recurs throughout this library and does the same work here as elsewhere. A single received quantity records what was accepted and destroys the evidence of a supplier discrepancy; two record both, and the difference is a supplier performance fact rather than a rounding error absorbed at the dock.

The received quantity is then decomposed into containers at the standard bin quantity, with a residual container for the remainder, and each container is given a unique barcode. Section 4.1 states the conservation constraint this decomposition must satisfy.

3.2Putaway, and the second key

The tagged container is carried to its allocated location. The operator scans the location barcode and the container barcode, and the system stores the container only when the pair agrees with the allocation.

The design decision here is that both keys are read from the physical world. A workflow that pre-selects the location on screen and asks the operator to confirm by scanning the container is checking one fact against a screen, not two facts against each other, and it fails silently whenever the screen and the aisle disagree.

3.3Issue under first-in-first-out, with shelf life alongside

Outward begins with a pick list from the host system. Material is picked from location under a first-in-first-out rule and issued to the line, with stock updated on issue.

Shelf life is carried per part — 750 days on the captured part record — and the interaction with first-in-first-out is worth being explicit about, because the two rules are not the same rule. First-in-first-out selects by receipt date; expiry-based selection selects by remaining life. They coincide only where every receipt of a part carries the same shelf life, which is the common case and not the general one.

The system reports first-in-first-out violations as a standing report rather than assuming its own enforcement worked. That is the correct posture for any enforcement mechanism operating through human hands: enforce, then verify that the enforcement held.

3.4The count window and in-flight movement

The count works best with a brief operational pause — four to eight hours for the initial physical pass, on the product's published figure. In-flight transactions across the count window are rolled back so that goods received or shipped during the count are excluded from the variance calculation.

This addresses the fourth error term of Section 1.1, and it is the one most often mishandled. A variance drawn against a book position that has moved since the count attributes post-count drift to shrinkage, and the attribution is invisible because both look like a quantity difference.

4.1Bin decomposition and its conservation constraint

A receipt of quantity Q is split into containers at the part's standard bin quantity, with one residual container holding the remainder.

For the captured receipt, 1,000 pieces at a standard bin quantity of 120 give n = 8 and r = 40: eight containers of 120 plus one of 40, which is exactly what the screen records.

The decomposition's real function is granularity. A count can verify only what is independently identified, so a receipt held as one undifferentiated quantity can be verified only in total, while nine tagged containers can be verified nine times. This is the same argument as the traceability tier structure elsewhere in this library: resolution is decided at capture and cannot be recovered afterwards.

4.2Two-key putaway and its error model

A putaway is accepted only when both scanned keys agree with the allocation. Stating this as a condition makes the residual error visible.

The correlation term is where designs of this class succeed or fail. Two keys read from two physical labels are close to independent, so a putaway error requires two simultaneous mistakes. A workflow where the screen pre-selects the location and only the container is scanned collapses rho, because a single wrong screen selection satisfies both conditions at once.

The control-against-warning distinction of Section 2.2 then determines what happens on a mismatch. Refusing to store means the container cannot enter the record in the wrong place; warning and permitting means it can, with a note nobody reads.

4.3Coverage: making double and missed counts impossible

The two dominant errors in a manual count are both coverage errors: a container counted more than once, and a container never reached. Both are eliminated by treating the count as a set operation over tags rather than a walk over aisles.

Two consequences follow that procedure alone cannot deliver. Double-counting is impossible rather than discouraged, because the second scan changes nothing. And the missed set is known while the counters are still on the floor, so flagged bins are re-counted before the team leaves rather than the following week from memory.

This is the mechanism behind the published reduction in count duration. The saving is not from scanning being faster than writing; it is from never running the count twice.

4.4The variance the reconciliation owes

Once the count closes, the system compares counted quantity to book quantity per container and reports the difference. How that difference is aggregated decides whether the report is evidence or reassurance.

The failure this prevents is specific and common. A facility with one location over by 400 pieces and another under by 400 has a net variance of zero and an absolute variance of 800. Reported by net alone, the count reconciles perfectly and the facility concludes its records are sound; in fact it has two errors, both of which will cause a pick failure, and neither of which it now knows about.

Record accuracy — the fraction of containers whose count matched exactly — is the figure a warehouse should manage against, because it is the probability that any given pick will find what the system promised. It is also the only one of the three that cannot be improved by compensating errors.

The variance report carries counter name, location, timestamp and before-and-after quantities per part, with discrepancies stratified by value impact — which is what makes an adjustment attributable rather than anonymous, and what a statutory audit asks for.

5.1Observed properties of the software

The last row is worth a note on honesty in both directions. The product page claims over fifteen reports; the captured menu holds nineteen. Where a count can be made it has been made here rather than repeating the marketing range.

5.2Figures published for this module

The count-duration figure is the one a buyer will weigh most heavily and the one carrying least evidence. It is a range with no facility size, no container population and no definition of what the phase includes. Section 4.3 gives the mechanism that would produce such a reduction — never running the count twice — which makes the claim structurally plausible without making it measured.

5.3Modelled effort avoided

The published return model for this product is stated as stock-take and reconciliation time saved. Its assumptions are printed here so a reader can substitute their own.

Two observations. The model's window is sixty days, not the count weekend — so what it prices is the reconciliation tail rather than the count itself, which is consistent with the product's own framing that the expensive part is the fortnight of spreadsheet archaeology afterwards. And it prices only labour: the shutdown hours of production and shipping are excluded, and in most facilities those dominate.

A buyer computing this properly should add the shutdown. If a facility loses forty-eight hours of production to a paper count and eight to a scanned one, the recovered output is likely to exceed the entire labour saving the published model reports.

6.1The annual count is a symptom

A facility that must halt for two days each year to discover what it holds is telling you that its records were not trusted for the preceding three hundred and sixty-three. The count is a consequence of that distrust, not a cause of accuracy.

This reframes what a tagging system is for. Its most valuable output is not the annual variance report but the record accuracy figure of Section 4.4, because that figure is what determines whether counts can eventually become sampled rather than exhaustive. A facility at 99.5% record accuracy can verify a stratified sample and defend it; one at 92% cannot, and will be counting everything for as long as that holds.

The controls of Sections 4.1 to 4.3 are what move that figure, and all three operate during the year rather than during the count. A system evaluated only on how fast it counts is being evaluated on the least important thing it does.

6.2Enforcement and verification are different jobs

The system enforces first-in-first-out at picking and separately reports violations of it. That is worth drawing out, because carrying both is unusual and correct.

Any enforcement that passes through human hands has a bypass. A picker facing an empty designated location and a full one behind it will take the accessible stock, and a well-designed system lets them — blocking the pick would stop the line. The violation report is what converts that necessary flexibility from an untracked leak into a measured exception with a location attached.

The same logic recommends caution with the shelf-life interaction of Section 3.3. Where receipts of a part carry different remaining lives, first-in-first-out and expiry-first selection diverge, and a system enforcing the first while an operator believes it enforces the second will issue expiring stock in strict receipt order.

6.3A capability reference framework for stock-take systems

6.4Generalisability

The coverage argument of Section 4.3 generalises to any exhaustive verification over an identified population: asset registers, tool cribs, library stock, calibration inventories. The absolute-against-net variance argument generalises further still, to every reconciliation where compensating errors are possible — which is most of them.

What does not generalise is the count-duration figure. It depends on facility size, container population, counter numbers and layout, none of which the published range states, and a facility adopting it as a plan would be planning against another site's geometry.

Appendix B.1Decomposition, and what it buys

A receipt of 1,000 pieces arrives for a part whose standard bin quantity is 120.

Applying Equation (decomp): n = floor(1000 / 120) = 8, r = 1000 — 960 = 40. Nine tagged containers — eight of 120 and one of 40 — and the identity closes: 8 x 120 + 40 = 1,000.

Now consider what each arrangement permits at count time. Held as one quantity, the receipt admits exactly one verification, which passes if the total is right however the stock is distributed. Held as nine containers, it admits nine, and a container in the wrong aisle produces a variance in two places rather than none.

Appendix B.2What the second key is worth

Suppose an operator scans the wrong location with probability 0.004 and picks up the wrong container with probability 0.006, over 20,000 putaways in a year.

With one key — container only, location pre-selected on screen — the undetected misplacement rate is 0.004, giving roughly 80 misplaced containers a year, each invisible until a pick fails.

With two independently read keys and rho near 1, Equation (putaway) gives 0.004 x 0.006 = 0.000024, or about one misplacement every two years at the same volume.

Now collapse the independence. If the screen pre-selects the location and the operator confirms by scanning whatever container is in hand, a single wrong screen selection satisfies both conditions, rho rises toward 1/p(c), and the undetected rate returns to roughly the single-key figure. The second scan happened; the second key did not.

Appendix B.3Three numbers from one count

A count covers 2,400 containers. 2,352 match exactly. Of the remaining 48: 22 are over by an average of 31 pieces at 18 currency units each, and 26 are under by an average of 27 pieces at 19 units each.

Applying Equation (variance): the positive value variance is 22 x 31 x 18 = 12,276 and the negative is 26 x 27 x 19 = 13,338. Net variance is -1,062. Absolute variance is 12,276 + 13,338 = 25,614. Record accuracy is 2,352 / 2,400 = 98.0%.

Reported by net alone the count looks excellent: a shortfall of 1,062 on a population of 2,400 containers is well inside any materiality threshold, and the ledger reconciles. Reported honestly, the facility has 48 wrong locations, an absolute error twenty-four times the net, and a two-percent probability that any given pick will not find what the system promised.